Discussion ChatGPT Can help writing Code for Orbiter

[N_Molson]

Nice RetroGlider lol...

 

[nbcfrosty]

AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms

New AI agent evolves algorithms for math and practical applications in computing by combining the creativity of large language models with automated evaluators

deepmind.google

AlphaEvolve discovered a simple yet remarkably effective heuristic to help Borg orchestrate Google's vast data centers more efficiently. This solution, now in production for over a year, continuously recovers, on average, 0.7% of Google’s worldwide compute resources.

Also:

Stanford CRFM

crfm.stanford.edu

GPU Kernel Scientist: An LLM-Driven Framework for Iterative Kernel Optimization

Optimizing GPU kernels for high performance is a complex task, often demanding deep architectural knowledge, extensive profiling, and iterative experimentation. This challenge is amplified when targeting newer or less-documented GPU architectures where traditional development aids are scarce...

arxiv.org

And ofc:

AlphaFold

AlphaFold has revealed millions of intricate 3D protein structures, and is helping scientists understand how all of life’s molecules interact.

deepmind.google

AlphaGeometry: An Olympiad-level AI system for geometry

Our AI system surpasses the state-of-the-art approach for geometry problems, advancing AI reasoning in mathematics

deepmind.google

 

[Buck Rogers]

OpenAI released 2 free "open source" models on Tuesday.
gpt-oss-120b and gpt-oss-20b push the frontier of open-weight reasoning models

 

[nbcfrosty]

 

[Buck Rogers]

Why OpenAI’s solution to AI hallucinations would kill ChatGPT tomorrow​

In short, the OpenAI paper inadvertently highlights an uncomfortable truth: the business incentives driving consumer AI development remain fundamentally misaligned with reducing hallucinations. Until these incentives change, hallucinations will persist.

Complete article at Converstion.

 

[jedidia]

It demonstrates that these aren’t just an unfortunate side effect of the way that AIs are currently trained, but are mathematically inevitable.

God dammit, I hate being right!

This revealed that nine benchmarks use binary grading systems that award zero points for AIs expressing uncertainty.


This creates what the authors term an “epidemic” of penalising honest responses. When an AI system says “I don’t know”, it receives the same score as giving completely wrong information.

In other words, we train them to behave like some rando on twitter. Great idea!

 

[DaveS]

God dammit, I hate being right!


In other words, we train them to behave like some rando on twitter. Great idea!

Anyone remember when MS tested their "AI" on Twitter? Didn't take long for it to lose any sanity it might have had at the beginning. Lead to MS having to pull it entirely.

 

[Urwumpe]

In other words, we train them to behave like some rando on twitter. Great idea!

Elon Musk constantly has to order fixes in Grok, because it opposes what he says....

 

[jedidia]

Anyone remember when MS tested their "AI" on Twitter? Didn't take long for it to lose any sanity it might have had at the beginning. Lead to MS having to pull it entirely.

A bit of another situation. MS didn't test... They thought it would be a good idea to train it by public conversation. Enter 4Chan, and before long it was denying the holocaust and hailing Hitler. It was a neat experiment clearly showing that training might need some oversight, an insight which any random parent on this planet could have given them for free...

 

[Urwumpe]

A bit of another situation. MS didn't test... They thought it would be a good idea to train it by public conversation. Enter 4Chan, and before long it was denying the holocaust and hailing Hitler. It was a neat experiment clearly showing that training might need some oversight, an insight which any random parent on this planet could have given them for free...

I disagree there, I think its: Most random parents.

Just have somebody in my social circles, who started to radicalize extremely after the Kirk shooting (despite being deep in the "unworthy" category of his worldview) and has the same media competence as Microsoft AI back then. And is parent.

Well, flags in the wind can swing all ways.

 

[steph]

Well, folks, for the second time in as many years, I have made the mistake of subscribing to GPT Plus (the first tier, not the pro).

Had a personal project that required writing some Java script, and,as my only coding experience involved Pascal way back in highschool, I asked the free version and , while it did write it, it didn't work, so I had the genius idea to upgrade.

Still didn't work, of course, although after a few failed attempts, it stopped providing code and just kept telling me it 'fixed it' and now it should work, apparently because it got into my browser and wrote rhe code itself. Now, that's something I'm actually glad it's not capable of doing (yet - there are apps it can bind to already, but user permission is required)

So, long story short, for about a month from now, if you have any ideas that I could pitch to the mighy GPT Plus, just let me know and I'll forward them, but I can't guarantee it won't be useless.

I think the problem is that, while it can do stories and all sorts of stuff, those are much more lenient towards changes/mistakes than code, where one wrong line means the damn thing doesn't work at all.

 

[JDat]

If you want something working (or good quality) make it yourself and don't hope or relay on others! And I mean human beings. If I see someone tell AI do things etc etc, I read it following: That, who asks question to AI, is more stupid than AI. And AI stupidity you already know by your own experience.

 

[nbcfrosty]

Well, folks, for the second time in as many years, I have made the mistake of subscribing to GPT Plus (the first tier, not the pro).

Had a personal project that required writing some Java script, and,as my only coding experience involved Pascal way back in highschool, I asked the free version and , while it did write it, it didn't work, so I had the genius idea to upgrade.

Still didn't work, of course, although after a few failed attempts, it stopped providing code and just kept telling me it 'fixed it' and now it should work, apparently because it got into my browser and wrote rhe code itself. Now, that's something I'm actually glad it's not capable of doing (yet - there are apps it can bind to already, but user permission is required)

So, long story short, for about a month from now, if you have any ideas that I could pitch to the mighy GPT Plus, just let me know and I'll forward them, but I can't guarantee it won't be useless.

I think the problem is that, while it can do stories and all sorts of stuff, those are much more lenient towards changes/mistakes than code, where one wrong line means the damn thing doesn't work at all.

you are using the wrong tool for the job.... What you want is a Claude Code subscription for $20/month or a Cursor sub or a MS Copilot sub.
ChatGpt is not good at coding, the models used there are not optimized for coding.

edit:
For coding stick to claude 4 sonnet

 

[N_Molson]

Actually I'm learning a lot with Gemini. It gives me code samples that (very often) work. And it teaches me new way to do things.

The key is to stay at a simple task level. Don't ask it to write your whole code, it doesn't have a 'global' vision. Also it tends to be decently good at orbital maths, and when you're not a maths person (like me), it helps.

 

[JDat]

Sometimes, when need to solve some problem while writing code, I ask question right into duckduckgo and pres AI help. It usually answers in a way that I can solve problem. Usually some casting error in C. Good thin, it also gives URL from where it took info, so I can read myself.

 

[GodAtum]

I paid for Claude Code and connected it to the Orbiter Gitlab. I asked it to convert it to DX12 but it failed completely. I tried something more simple to move TransX to an external Java program but read data from the running sim but that didn't really work either

 

[N_Molson]

You're asking the impossible. Again, AI will not code an entire project for you. It is way beyond its real abilities, whatever the advertisment says. Don't waste money on this.

What it can do is write something of the scale of a function. But you still have to be the Architect, and you need to clean behind the AI (so you have to know the Orbiter API and how it works).

 

[nbcfrosty]

You're asking the impossible. Again, AI will not code an entire project for you. It is way beyond its real abilities, whatever the advertisment says. Don't waste money on this.

What it can do is write something of the scale of a function. But you still have to be the Architect, and you need to clean behind the AI (so you have to know the Orbiter API and how it works).

That’s not remotely true anymore. There is a whole bunch of software between porting orbiter to use dx12 autonomously one shot and just a function.

You can look in this thread where 3 mfds were coded using ai. I create projects using cursor and Claude code all the time.

 

[nbcfrosty]

I paid for Claude Code and connected it to the Orbiter Gitlab. I asked it to convert it to DX12 but it failed completely. I tried something more simple to move TransX to an external Java program but read data from the running sim but that didn't really work either

What you are asking for are not simple things…

 

[GodAtum]

That’s not remotely true anymore. There is a whole bunch of software between porting orbiter to use dx12 autonomously one shot and just a function.

You can look in this thread where 3 mfds were coded using ai. I create projects using cursor and Claude code all the time.

Would you mind giving some of your promts for Claude Code?